Comment: DeepSeek can shake up the AI sector

These are times when GenAI breakthroughs have become so routine to not even qualify as news. Yet, DeepSeek ruffled financial markets in the past few days. News relating to the latest GenAI kid on the block was sufficient to see the world’s leading chipmaker’s shares plummet by around one-fifth in a single day. The financial tremors were felt wider across the tech industry. DeepSeek’s GenAI model - R1 - released in Nov 2024 has been noted to compete well with popular models, but it is definitely not anything like a step-change. That’s when you look at the technology. In contrast, when you look at other things - such as operational and business models - the differences are stark. DeepSeek is a game-changer in several ways. 

The AI sector, it has been noted, is dominated by industry in ways that no other sector is. Academia is increasingly playing second fiddle. This is at least due to two factors on which the academia finds it hard to compete with the industry: the dependency of AI on massive computational resources and very large datasets. These factors have led to concentration of AI among ‘big tech’ in ways that smaller firms increasingly find it hard to even sustain. The trend towards monopolisation started during the age of information services (e.g. Google), continued throughout the social media (e.g. Meta) and gig economy (e.g. Uber) sectors, and has shown signs of intensification in the GenAI sector (e.g. OpenAI). It is this backdrop that makes DeepSeek important.

Open Source

First, DeepSeek’s models are all open source. The blueprint to train the models is available (with its open implementations being developed), and so are the trained models. Notably though, the data to train the models are not yet open. One can download the models and run on their own computer, without an internet connection. If you need to train your own R1 on your private dataset, you can do so too - as long as you have appropriate computing resources. These are massive differences that contrast well with the non-availability of either training code or blueprints or models for popular contemporary models such as ChatGPT. In fact, when you use ChatGPT, it knows that you are using it - when you pause an extra second or express disagreement, you are contributing data to further improve ChatGPT. You can’t use ChatGPT as a product in the traditional sense; even if you are a paid user, you can only use it under constant algorithmic oversight. DeepSeek is not the first open-source large language model (LLM) - beyond an early wave of enthusiasm among big-tech (e.g., Facebook’s Llama) in open-sourcing models, there are thriving initiatives in that space, such as EleutherAI. Yet, an open-source model that competes well with commercial models is significant - it breathes fresh energy into the space of open-source AI. 

Reduced Computation

Second, DeepSeek has been noted to pose significantly reduced computational requirements for training its models. This is in stark contrast with an exponentially increasing computational requirement that AI has been notorious for. It is also notable that massive compute requirements yield well to monopolisation of the space, since it increases the entry barrier for small firms. 

The big tech giants, riding high on computational resources and market share, have been lukewarm towards funding research into reducing computational expenses for AI. The government funding landscape and the general community had started to accept that AI, by nature, is a power-guzzler. Against this backdrop, DeepSeek’s massive advances in reducing computational training requirements for AI should open up a brave new world of greener AI - a field in which smaller players can meaningfully compete. 

Non-financial Motives

Third, possibly the most surprising of all, DeepSeek has shown disinterest in spawning significant commercial activity out of their AI models. The company is led by a tech idealist CEO who has not been hesitant to say - in a 2024 interview - that he would put ‘right and wrong’ before ‘profits and losses’ and that he would always be against exorbitant profits. He appears to put his weight behind delivering original innovation, while slyly attacking the current paradigm within the AI world as one of imitations. This throws light into the kind of technology products and pricing model DeepSeek may pursue, and the kind of price wars it would trigger in a corporate dominated space. A player who doesn’t put financial motives first isn’t the kind of player that Silicon Valley is used to - and the outcome is anything but a foregone conclusion. 

Winds of Change

The above aspects indicate potential winds of change. It may be useful to remind ourselves of a story from the past where a paradigm shift upended a dominant business model. Let’s go back to the 1990s when owning a PC almost implied being a Microsoft Windows customer. What came out of the blue then was the free and open-source movement, which delivered a free operating system, Linux. Across the world, and in particular in the Global South where Windows was felt as very expensive, Linux spread like wildfire. Even today, significant sectors run on free and open-source software solutions of various kinds. Since then, big corporate power has clawed into the open-source sector through devising various business models, but the fact remains that the free software movement is still backed by a community of passionate tech enthusiasts. Could DeepSeek similarly upend the corporate domination in GenAI tech or is this just a blip? Only time will tell. 

A Word of Caution

While DeepSeek has room for optimism, we ought to be cautious, very cautious. What DeepSeek will do in the short run will become clearer over the next few days. For one, it will make AI much more pervasive, and way cheaper. It is highly unlikely that ChatGPT will continue to have significant takers for its ‘Pro’ plan unless it brings down the current price tag of $200pm significantly. The pervasive availability of AI will also intensify fears about job losses. More people may feel threatened by it within their jobs. Those who don’t use AI will find it harder to stay relevant in their jobs. Whether or not such fears are well-founded, tech bosses may find it easier to push down wages in a climate of accentuated job insecurity. 

Stepping back a bit, if what we seek are progressive changes in the tech industry, we ought to reorient it from its profiteering obsession towards ways that it can meaningfully address the needs of society. However, in the past few years, the AI sector - as noted earlier - has arguably witnessed changes that intensify the problems through deepening monopolisation tendencies, with  big-tech growing at the expense of everything else. Whatever DeepSeek is, it is clear it stands against the flow. Will the new kid on the block herald winds of change? We mustn't lose hope. 

Deepak P is Associate Professor in AI at Queen’s University Belfast